Quantum Counters

advertisement
Quantum Counters
Smita Krishnaswamy
Igor L. Markov
John P. Hayes
Contents
Motivation
Background
Previous work
Basic Circuit
Good Parameters
Improved Circuits
Current Work /Conclusions
BACKGROUND
Motivation
Goal of QC thus far has been to solve
problems faster than classical computing
Deutsch’s algorithm first example of
algorithm with faster quantum algorithm
Our goal is to obtain exponential memory
improvement for a specific problem

We use sequential circuits to achieve this
Sequential Circuits
Sequential circuits contain a combinational
portion and a memory portion
Combinational portion is re-used and
therefore usually simpler
Sequential circuits are modeled by finite
automata
Finite Automata
5-tuple {Q, ,  , q0 , Qacc }
Q=set of states
Σ=input alphabet
q0=starting state
Qacc =set of accepting states
 = transition function
 :Q  Q
More Finite Automata
The memory portion stores state info
An FSA for a counter that counts to 4:
Quantum Finite Automata
RFA: Reversible finite automata i.e. only
one arrow going into each state
A QFA is a reversible finite automata that
transitions between quantum states
Q=the vector space of state vectors
Qacc=the accepting subspace with an
operator Pacc that projects onto it
δ is a unitary matrix
Example
RFA for 2-counter (can do it in 2 states)
RFA’s cannot recognize E={aj|j=2k+3}
PREVIOUS WORK
Prime Counter
For p prime let language Lp  {a : p | j}
Any deterministic FA recognizing Lp takes
at least p states
Ambainis and Freivalds [1] show that a
QFA needs only O(log p) states
O(log p) states requires only O(log log p)
qubits. This is an exponential decrease
j
QFA for L p
Set of states : Q = {|0>,|1>}
Starting state: q0=|0>
Set of accepting states: Qacc = span(|0>)
Next state function δ:
 cos( ) i sin( )
i sin( ) cos( ) 


Counter QFA
Pick a rotation angle Φ=2πk/p, 0<k<p
For input aj , rotate qubit j times by Φ
If j is a multiple of p then the state is
definitely |0>, else we have |1> with
probability cos2(Φ)
Want to pick a set of k’s that increases the
probability of obtaining state |1> for every j
1-qubit Counter for p=5
j= 1 mod 5
j=2 mod 5
Φ
j = 0 mod 5
j=3 mod 5
j=4 mod 5
Sequence of QFAs
This QFA rejects any x not in Lp with varying
probability of error ranging from 0 to 1
Therefore any one of these QFAs is not enough
We can pick a sequence of 8 ln p QFA’S with
different values for k where Ф =2 π k/ p
The values for k can be picked such that the
probability of error is always less than 7/8
Proof Sketch
• At least half of all of the k’s that we
consider reject any given aj not in Lp with
probability at least ½
• There is a sequence of length 8 ln p that ¼
of all elements reject every aj not in Lp with
probability ½ (This follows from Chernoff
Bounds)
Problems To Address
No explicit circuit construction given
No explicit description of the sequence of
angle parameters given
Loose error estimate
CIRCUIT
IMPLEMENTATION
Quantum Circuits
Quantum operators are unitary matrices
A larger matrix is broken into a matrix or
tensor product of 2x2 matrices (1-qubit
gates)
Gate library: {Rx, Ry, Rz ,C-NOT, NOT}
Rx (Φ )=  cos( ) i sin( )
i sin( ) cos( ) 


Gate Library
Ry (Φ )=
Rz (Φ )=
NOT =
cos( )  sin( )
 sin( ) cos( ) 


i
e

 0
0
1

0 
 i 
e 
1
0

Gate Library
C-NOT =
1
0

0

0
0
1
0
0
0
0
0
1
0
0
1

0
K-controlled Rotations
H=Ry(π/4)
Controlled Rotations
Barenco et. al. [2] give construction for k-controlled
rotations from basic gates
They are constructed from K-controlled NOT gates
and
1-controlled rotation gates
Implementation
The QFAs cannot be implemented
separately without wasting space

Need O(log p) qubits for this
Can implement as a block-diagonal matrix
Each block on the diagonal is an Rx
Can be implemented with controlled
rotations
Block-Diagonal Matrix
 2k1

0
0
0
 Rx ( p )



2k 2


0
Rx (
)
0
0


p


2k3
0
0
Rx (
)
0


p


2k 4 

0
0
0
R
(
)
x


p


Basic Circuit
Each block diagonal corresponds to one
k-controlled rotation gate
Example:
GOOD PARAMETERS
Error Probability
The probability of error for this circuit is:
Perr  max
1 j  p
1
2 2ki j
cos (
)

n i 1
p
n
The expression is the sum over the error
contribution of each ki
Try to minimize the maximum error for any value
of j<p
Picking parameters:Observations
• Theorem: No parameter set can have
probability of error <½
• Proof:
For a given k the average probability of error
over all j’s is ½
 A sequence of k’s has the same average
probability of error
 Therefore Perr which is the max of all of these
has to be > ½

Rejection Patterns
• Each parameter is “good” for a ½ of the j’s
and they occur in a specific pattern
Greedy Selection of Parameters
Try to cover as much new area as possible
Continue this process until all parameters
are rejected with a certain probability
Obtain a set of parameters that follow the
sequence milj where m and l are constants
Use mutually prime values of m and l to
avoid repetition

Usually m=2 and l=3
Asymptotic Behavior
These params give low Perr for all p’s
Estimating Error Bounds : Idea
Discretize the cosine expression
If sin2(φ)> ½ regard it as 1/2
If sin2(φ)< ½ regard it as 0
The area covered by a parameter k : the portion of
the unit circle where sin2 (2πk/p) >½
For the parameters milj can get recursive
expressions for which areas are covered by n of
the k’s
If n was half the total number of k’s.The
probability lower bounded by (1/2)(1/2)=(1/4)
IMPROVED CIRCUITS
Circuit Complexity
Basic block-diagonal circuit: too many gates
There are only O(log log p) qubits where
as there are O(log p) gates
Different circuit decomposition may yield
better results

Some reductions are possible
Reductions
Order the parameters such that their
controls are in Gray Code order
Only one C-NOT is required between any
set of C-NOTS
Reducing Control Bits
Can use only log p different rotations
We apply them with fewer control bits
by using binary addition
Other Techniques
Random circuit simulations


Pick a set of k-controlled circuits repeatedly
Save the best
Greedy Simulation
Pick gates one by one first go through
controls then through rotations, p2
possibilities
Continue this while probability decreases
Order does not matter
Diagonalization
 2k1

R
(
)
 z p

H
H


2

k
1


Rz (
)



H
H


p






2k1




H
H
R
(
)

z



p
H
H






2k1 

R
(
)
z

p 

Using Diagonal Computations
Diagonal computation uses identity
HRx(θ)H= Rz(θ)
NDIAG algorithm by Bullock and Markov [4]
reduces gate counts by tensor-product decomposition
This technique may not be useful for this circuit for
an inherent reason



the range of the parameters too big
Tensor product of two rotations adds the angles.
Need to explore this: could mean there are NO good
circuits for this computation.
Tensor Products
Diagonal unitary matrices
have the form
Tensor products of two
such matrices
have the form
 e i

0
0
i 
e 
ei (  )



i (  )
e




e i (   )

i (   ) 
e


On-going Work: Proof Sketch
We want a circuit with O(log log p) gates and we
are trying to combine them to form a computation
that has O(log p) rotations.
This is possible if we use the gates as rotations
with angles that add like binary numbers.
Problem: We want O(log p) rotations spread out in
the range [1,p]. Using O(log log p) rotations we
can either get [1, log p] consecutive rotation
angles or [1 p] rotation angles with big holes in
between.
Current Work
Finding better circuits
Finding circuits or proving that no good
circuits exist for the greedy parameters
Coming up with an analytical error bound
for these parameters
Empirically the error value is around .60
Conclusions
Studied counters with exp memory savings
Can construct unitary computations
 Can construct quantum circuits

Open Questions
Are there any polynomial sized circuits or is
there a size-accuracy tradeoff?
Will Fourier transforms or other techniques
give friendlier parameters?
Do other quantum sequential circuits
improve memory usage over classical
circuits?
References
Download